What are barren plateaus?
Regions of a variational circuit's parameter space where gradients vanish exponentially with qubit count, making training practically impossible.
McClean and colleagues (2018) showed that in sufficiently deep random parameterised circuits, the variances of the gradients are exponentially small in the number of qubits. This means that the gradient at a random starting point is very close to zero, and estimating it accurately enough requires an exponential number of measurements.
Factors linked to the phenomenon: circuit depth and expressivity, Global Cost functions that measure all qubits, excessive entanglement, and noise. Proposed strategies include local cost functions, shallow or problem-tailored circuits, and smart parameter initialisation; there is no general solution.
In the lab you can explore the loss function's “landscape” in the landscape experiment, but the small sizes possible in the browser are not enough to see the phenomenon at full strength.
Related topics
References
- Jarrod R. McClean; Sergio Boixo; Vadim N. Smelyanskiy; Ryan Babbush; Hartmut Neven (2018). Barren plateaus in quantum neural network training landscapes. Nature Communications 9:4812 doi:10.1038/s41467-018-07090-4
- M. Cerezo et al. (2021). Variational quantum algorithms. Nature Reviews Physics 3:625–644 doi:10.1038/s42254-021-00348-9
